Sei sulla pagina 1di 12

ECE 361: Fundamentals of Digital Communications

Lecture 22: Time Diversity

Introduction
We have seen that the communication (even with coding) over a slow fading flat wireless
channel has very poor reliability. This is because there is a significant probability that the
channel is in outage and this event dominates the total error event (which also include the
effect of additive noise). The key to improving the reliability of reception (and this is really
required for wireless systems to work as desired in practice) is to reduce the chance that the
channel is in outage. This is done by communicating over different channels and the name
of the game is to harness the diversity available in the wireless channel. There are three
main types: temporal, frequency and antennas. We will see each one over separate lectures,
starting time diversity.

Time Diversity Channel


The temporal diversity occurs in a wireless channel due to mobility. The idea is to code and
communicate over the order of time over which the channel changes (called the coherence
time). This means that we need enough delay tolerance by the data being communicated.
A simple single-tap model that captures time diversity is the following:

y = h x + w , = 1, . . . , L. (1)

Here the index represents a single time sample over different coherence time intervals. By
interleaving across different coherence time intervals, we can get to the time diversity channel
in Equation (1). A good statistical model for the channel coefficients h1 , . . . , hL is that they
are all independent and identically distributed. Further more, in a environment with lots of
multipath we can suppose that they are complex Gaussian random variables (the so-called
Rayleigh fading).

A Single Bit Over a Time Diversity Channel


For ease of notation, we start out with L = 2 and just a single bit to transmit. The simplest
way to do this is to repeat the same symbol: i.e., we set

x1 = x2 = x E. (2)

1
At the receiver, we have four (real) voltages:
[y1 ] = [h1 ] x + [w1 ] , (3)
[y1 ] = [h1 ] x + [w1 ] , (4)
[y2 ] = [h2 ] x + [w2 ] , (5)
[y2 ] = [h2 ] x + [w2 ] . (6)
As usual, we suppose coherent reception, i.e., the receiver has full knowledge of the exact
channel coefficients h1 , h2 . By now, it is quite clear that the receiver might as well take the
appropriate weighted linear combination to generate a single real voltage and then make the
decision with respect to the single transmit voltage x. This is the matched filter operation:
y MF = [y1 ] [h1 ] + [y1 ] [h1 ] + [y2 ] [h2 ] + [y2 ] [h2 ] . (7)
Using the complex number notation,
y MF = [h1 y1 ] + [h2 y2 ] , (8)
= |h1 |2 + |h2 |2 x + w.

(9)
Here w is a real Gaussian random variable because it is the sum of four independent and
identically distributed Gaussians, but they get scaled by the real and imaginary parts of h1
and h2 . Therefore w is zero mean and has variance:
2
|h1 |2 + |h2 |2 .

Var(w)
= (10)
2
The decision rule is now simply the nearest neighbor rule:

decide x = + E if y MF > 0, (11)
and zero, otherwise.

Probability of Error
The dynamic error probability is now readily calculated (since the channel cf. Equation (9)
is now of an AWGN type):

2 2

(|h1 | + |h2 | ) E
Pedynamic = Q p . (12)
|h 1 |2 + |h |2
2
2

From the transmitters perspective, it makes sense to calculate the average error probability,
averaged over the statistics of the two channels h1 , h2 :
h p i
Pe = E Q 2 2
2SNR (|h1 | + |h2 | ) . (13)

2
For the statistical model we suggested earlier (independent Rayleigh fading), this expression
can be evaluated exactly: 2
1
Pe = (1 + ) (14)
2
where r
def ASNR
= . (15)
1 + ASNR
Much like we did in earlier lectures, we can try to simplify Equation (14) in order to better
see the relationship between unreliability level and SN R. Specifically, at high SNR (i.e.,
SN R >> 1), we have
1
(1 ) (16)
ASNR
(1 + ) 2. (17)

Substituting this approximations into Equation (14), we conclude that


1
Pe = . (18)
8A2 SNR2

Diversity Gain
Equation (18) that captures the approximate relationship between transmit SNR and un-
reliability level is instructive: we see that a doubling in SNR reduces the unreliability by a
quarter factor. This is significant improvement when compared to the earlier case of no di-
versity: in that case, a doubling of SNR only decreased Pe by a factor of half. The exponent
of SNR in Equation (18) is referred to as diversity gain, and in this case it is equal to 2.

Generalization
We have worked with L = 2 for simplicity. For a general L, we have the natural extension
to repetition coding at the transmitter and matched filter reception (with nearest neighbor
decoding) at the receiver. Equation (9) generalizes to
L !
X
MF 2
y = |h | x + w, (19)
=1

where w
is real Gaussian with zero mean and variance (cf. Equation (10))
L !
X
2 2
|h | . (20)
=1
2

3
Thus, Equation (13) generalizes to
vu L !
u X
Pe = E Q t2SN R |h |2 . (21)
=1

Again, there is an exact expression for the unreliability level when the channel coefficients
are independent Rayleigh distributed (as a generalization of Equation (14)):
L1
L X k
(1 ) L1+k 1+
, (22)
2 k 2
k=0

where is as before (cf. Equation (15)). Finally, we can look for a high SNR approximation.
Equation (18) now becomes

2L 1 1
Pe . (23)
L (4ASNR)L

The diversity gain is now L: doubling of SNR reduces the unreliability by a factor of
1
. (24)
2L
Figure 1 depicts the power of having time diversity. It is instructive to see that to even
have a reliably high unreliability level (per bit) of105 we need a SNR of almost 40dB with
first order diversity, but this reduces to just 12 dB with fifth order diversity. The diversity
gain is even starker at lower unreliability levels.

Looking Ahead
In the next two lectures we will look at two other modes of diversity: frequency and antennas.
These come in handly particularly when time diversity is either absent or cannot be effective
harnessed due to the tight delay constraint of the data (say, gaming applications).

4
Figure 1: Bit error probability as a function of different diversity orders L.

5
ECE 361: Fundamentals of Digital Communications
Lecture 23: Frequency Diversity

Introduction
In many instances, temporal diversity is either not available (in stationary scenarios) or
cannot be efficiently harnessed due to the strict delay constraints of the data being commu-
nicated. In these instances, and as an added source of diversity, looking to the frequency
domain is a natural option.

Frequency Diversity Channel


Frequency diversity occurs in channels where the multipaths are spread out far enough,
relative to the sampling period, so that multiple copies of the same transmit symbol are
received over different received samples. Basically, we want a multitap ISI channel response:
the L-tap wireless channel
L1
X
y[m] = h x[m ] + w[m], m1 (1)
=0

is said to have L-fold frequency diversity. The diversity option comes about because the
different channel taps h0 , . . . , hL1 are the result of different multipath combinations and
are appropriately modeled as statistically independent. Thus the same transmit symbol (say
x[m]) gets received multiple times, each over a statistically independent channel, (in this
case, at times m, m + 1, . . . , m + L 1).

A Single Bit Over a Frequency Diversity Channel


Suppose we have just a singlebit to transmit. The simplest way to do this is to transmit
the appropriate symbol x = E and stay silent. : i.e., we set

x[1] = E, (2)
x[m] = 0 m > 1. (3)

At the receiver, we have L (complex) voltages that all contain the same transmit symbol
immersed in multiplicative and additive noises:

y[ + 1] = h x + w[ + 1], = 0, . . . , L 1. (4)

As before, we suppose coherent reception, i.e., the receiver has full knowledge (due to accurate
channel tracking) of the channel coefficients h0 , . . . , hL1 . We see that the situation is entirely

1
similar to that of the time diversity channel with repetition coding (as seen in the previous
lecture). Then, the appropriate strategy at the receiver (as seen several times, including in
the previous lecture on time diversity) is to match filter:
"L+1 #
def
X
y MF = h y[ + 1] (5)
=0
L1 !
X
= |h |2 x + w.
(6)
=0

Diversity Gain
Here w
is real Gaussian with zero mean and variance
L !
X
2 2
|h | . (7)
=1
2

Thus, the average error probability is generalizes to


v u L1 !
u X
Pe = E Q t2SNR |h |2 . (8)
=0

Again, there is an exact expression for the unreliability level when the channel coefficients
are independent Rayleigh distributed (just as in the time diversity case):
L1
L X k
(1 ) L1+k 1+
, (9)
2 k 2
k=0

where is as in the previous lecture. Finally, we can look for a high SNR approximation;
as earlier, we have:
2L 1 1
Pe . (10)
L (4ASNR)L
The diversity gain is now L: doubling of SNR reduces the unreliability by a factor of
1
. (11)
2L

2
OFDM
The approach so far has been very simple: by being silent over L 1 successive symbols
for every single symbol transmission we have converted the frequency diversity channel into
a time diversity one with repetition coding. While this allowed for easy analysis (we could
readily borrow from our earlier calculations) and easy harnessing of frequency diversity (so
reliability really improved) it has a glaring drawback: we are only transmitting once every
L symbols and this is a serious reduction in data rate. If we choose not to stay silent, then
we have to deal with ISI while still trying to harness frequency diversity. A natural way to
do this is to use OFDM. This way, we convert (at an OFDM block level) to the channel:
n xn + wn , n = 0, . . . , Nc 1.
yn = h (12)
This looks very much like the time diversity channel, except that the index n represents the
tones or sub-channels of OFDM.
It is interesting to ask what diversity gain is possible from this channel. Well, since the
channel really is the one we started out with (cf. Equation (1)), no more than L-order fre-
quency diversity gain should be possible (there are only L independent channels!). However,
a cursory look at Equation (12) suggests that there might be Nc order diversity gain, one
for each sub-channel. Clearly this cannot be correct since Nc is typically much larger than
L in OFDM.
The catch is that the OFDM channel coefficients h 0, . . . , h
Nc 1 are not statistically inde-
pendent. They are the discrete Fourier transform output of the input vector
[h0 , . . . , hL1 , 0, 0, . . . , 0] . (13)
Thus there are only about L truly independent OFDM channel coefficients and the di-
versity gain is restricted. Roughly speaking, once in every Nc /L tones fades independently.
Since the bandwidth of each tone is about W1 (where W is the total bandwidth of commu-
nication) we can say that the random discrete time frequency response H(f ) is independent
over frequencies that are apart by about W/L. This separation of frequency over which the
channel response is statistically independent from one another is known as the coherence
bandwidth (analogous to the coherence time we have seen earlier).
The coherence bandwidth restricts the total amount of frequency diversity available to
harness. Physically, this is restricted by the delay spread and communication bandwidth
(which decides the sampling rate). Most practical wireless systems harness frequency diver-
sity by a combination of having wide enough bandwidth of communication or a narrowband
of communication but one that is hopped over different carrier frequencies.

Looking Ahead
In the next lecture we will see yet another form of diversity: spatial. It is harnessed by
having antennas, at both transmitter and receiver.

3
ECE 361: Fundamentals of Digital Communications
Lecture 24: Antenna Diversity

Introduction
Time diversity comes about from mobility, which may not always be present (at least over the
time scale of the application data being communicated). Frequency diversity comes about
from good delay spread (relative to the bandwidth of communication) and may not always be
available. Fortunately, there is another mode of diversity, spatial, that can be harnessed even
under stationary scenarios with narrow delay spread. Antennas, transmit and/or receive,
are a natural element to capture spatial diversity. Antenna diversity is a form of time
diversity because instead of a single antenna moving in time we have multiple antennas
capturing the same mobility by being in different points in space. This means that antenna
and time diversity do not really occur independent of each other (while time and frequency
diversity occur quite independent of each other).
Antennas come in two forms: receive and transmit. Both forms provide diversity gain,
but the nature of harnessing them are quite different. Getting a better understanding of
how this is done is the focus of this lecture. We will also get to see the differences between
transmit and receive antennas. To be able to focus on the new aspect antennas bring into
the picture, our channel models feature no time and frequency diversity.

Receive Antenna Channel


The single transmit and multiple receive antenna channel is the following (with no time and
frequency diversity):

y [m] = h x[m] + w [m], = 1, . . . , L, m 1. (1)

Here L is the number of received antennas. Coherent reception is supposed, as usual. We


observe that the channel in Equation (1) is identical to the transmit diversity one with
repetition coding. So, the best strategy for the receiver is to match filter the L voltages:
" L #
def
X
y MF [m] = h y [m] , (2)
=1
L
!
X
= |h |2 x[m] + w[m].
(3)
1

1
As earlier, w[m]
is real noise voltage with Gaussian statistics: zero mean and variance equal
to L !
X 2
|h |2 . (4)
1
2

Even though the scenario looks eerily like time diversity with repetition coding (indeed, the
calculations are rather identical), there are two key differences:

the repetition was done by nature in the spatial dimension. So, no time was wasted
in achieving the repetition. This means that there is no corresponding loss in data
rate.

since there was no actual repetition at the transmitter, there is none of the correspond-
ing loss of transmit power. In other words, we receive more energy than we did in the
single antenna setting. Of course, this model makes sense only when we have a small
number of antennas and the energy received per antenna is a small fraction of the
total transmit energy. Indeed, this model will break down if we have a huge number
of antennas: surely, we cannot receive more signal energy than transmitted.

If the antennas are spaced reasonably apart (say about a wavelength distance from each
other), then a rich multipath environment will result in statistically independent channel
coefficients h1 , . . . , hL . This means that we have the L fold diversity gain as earlier. It is
important to bear in mind that this diversity gain is in addition to the L fold power gain that
was obtained due to the L receive antennas (these resulted in an L fold increase in receive
energy, as compared with the single antenna situation).
To summarize:

receive antennas are just an unalloyed good. They provide both power and
diversity gains.

Transmit Antenna Channel


The multiple transmit, single receive antenna channel (with no time and frequency diversity)
is modeled as:
XL
y[m] = h x [m] + w[m], m 1. (5)
=1

In other words, L distinct transmit (complex) voltages result in a single receive (complex)
voltage. Again, if the antennas are spaced far enough apart (say a distance of one wavelength)
in a good multipath environment then an appropriate statistical model would be to consider
h1 , . . . , hL to be statistically independent (and complex Gaussian distributed).

2
How does one harness the transmit diversity gain? Naive repetition coding, transmitting
the same symbol over each antenna, is not going to work. To see this, suppose
x[m]
x [m] = , = 1, . . . , L. (6)
L
Here we have normalized by L to have the same total transmit power as if we had a single
transmit antenna. Then the received signal is
L !
X h
y[m] = x[m] + w[m]. (7)
=1
L
The effective channel L !
X h
(8)
=1
L
is also Rayleigh fading: a sum of independent complex Gaussian random variables is still a
complex Gaussian random variable!
But a small twist will serve to fix the problem: we can repeat over the antennas, but use
only one at a time. Consider the following strategy:

x if = m
x [m] = (9)
0 else.
Then the first L received symbols are
y[] = h x + w[], = 1, . . . , L. (10)
This scenario is now identical to that of the receive antenna channel (cf. Equation (1)).
So, the diversity gain is readily harnessed (by matched filtering at the receiver). But two
important differences are worth highlighting:
there is no power gain from transmit antennas (this is because we actually transmitted
L times more energy than in the receive antenna scenario);
there is a loss in data rate by a factor of 1/L (this is because we actually used the
transmit one antenna at a time).
Modern communication techniques offer ways of using all the transmit antennas simul-
taneously while still allowing the harnessing of transmit diversity. This is the area of space
time coding. While this is a bit beyond the scope of this lecture, we can point out a simple
scheme that harnesses transmit diversity without sacrificing the data rate entailed by using
one antenna at a time. Consider the following transmission scheme:
x [m] = x1 [m 1], > 1, m > 2. (11)

3
With this transmission technique, the received signal is
L
X
y[m] = h x[m 1] + w[m]. (12)
=1

In other words we have converted the transmit antenna channel into one with frequency
diversity. Now a technique such as OFDM will allow the harnessing of the available diversity
gain. This technique, known as delay diversity, was one of the early space time coding
techniques.

Looking Ahead
All wireless systems harness one or more forms of diversity. This is the key to improving
the reliability of wireless communication. In the next (and final) lecture, we will see the key
new challenges (beyond those that arise in reliable point to point wireless communication)
in building a wireless system where there are many users communication at the same time.
We study these challenges in the context of the cellular wireless system.

Potrebbero piacerti anche